Dmitry Mikushin
2012-Dec-31 17:40 UTC
[LLVMdev] [DragonEgg] [Polly] Should we expect DragonEgg to produce identical LLVM IR for identical GIMPLE?
Dear all,
In our compiler we use a modified version LLVM Polly, which is very
sensitive to proper code generation. Among the number of limitations, the
loop region (enclosed by phi node on induction variable and branch) is
required to be free of additional memory-dependent branches. In other
words, there must be no conditional "br" instructions below phi nodes.
The
problem we are facing is that from *identical* GIMPLE for 3d loop used in
different contexts DragonEgg may generate LLVM IR either conforming the
described limitation, or violating it.
Let's consider two examples. In first one some simple 3D loop is used
directly in main program. In second - the same 3D loop is extracted into
separate subroutine called from main. Attached source code and listings for
GIMPLE and LLVM IR show that although GIMPLE codes are similar, by some
reason in first case branching goes under phi nodes, making Polly to fail
with parallelizing the resion. In second case everything is fine.
I have not looked into DragonEgg internals enough yet, but before I'll do,
the question is: should we expect DragonEgg to produce identical LLVM IRs
for identical GIMPLEs? The case shown here is really really important. The
success of parallelization utilities that are currently in a quite good
shape (thanks to their developers!) nowadays heavily relies on the code
quality: if IR is too rough, we can't do much about it :(
Many thanks and Happy New Year!
- Dima.
1) Bad LLVM IR:
==============
marcusmae at M17xR4:~/forge/kernelgen/tests/behavior/demo_f$
KERNELGEN_FALLBACK=1 kernelgen-gfortran -fplugin=dragonegg.so
-fplugin-arg-dragonegg-emit-ir -fplugin-arg-dragonegg-llvm-ir-optimize=0 -S
-c demo_f.f90 -o - | opt -O3 -S -o -
"161.i": ; preds =
%"160.i",
%"159.i"
call void bitcast (void (...)* @_gfortran_cpu_time_4 to void
(float*)*)(float* %start.i) nounwind
%204 = load i32* %ns.i, align 4
%205 = icmp sgt i32 %204, 0
br i1 %205, label %"162.preheader.i", label %"170.i"
"162.preheader.i": ; preds =
%"161.i"
%206 = bitcast i8* %x.0.0.i to float*
%207 = add i64 %y.3.2.0.0.i, %y.3.1.0.0.i
%208 = bitcast i8* %142 to float*
%.pre.i = load i32* %ny.i, align 4
%209 = icmp sgt i32 %.pre.i, 0
br label %"162.i"
"162.i": ; preds =
%"168.i",
%"162.preheader.i"
%210 = phi i32 [ %240, %"168.i" ], [ 1, %"162.preheader.i"
]
; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ THIS BRANCH CREATES
A PROBLEM
br i1 %209, label %"163.preheader.i", label %"168.i"
; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ THIS BRANCH CREATES
A PROBLEM
"163.preheader.i": ; preds =
%"162.i"
%211 = sext i32 %210 to i64
%212 = mul i64 %211, %86
%213 = mul i64 %211, %207
%.pre118.i = load i32* %nx.i, align 4
%214 = icmp sgt i32 %.pre118.i, 0
%215 = add i64 %212, %98
%216 = add i64 %213, %y.1.0.i
br label %"163.i"
"163.i": ; preds =
%"166.i",
%"163.preheader.i"
%217 = phi i32 [ %238, %"166.i" ], [ 1, %"163.preheader.i"
]
; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ THIS BRANCH CREATES
A PROBLEM
br i1 %214, label %"164.preheader.i", label %"166.i"
; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ THIS BRANCH CREATES
A PROBLEM
"164.preheader.i": ; preds =
%"163.i"
%218 = sext i32 %217 to i64
%219 = mul i64 %218, %72
%220 = add i64 %215, %219
br label %"164.i"
"164.i": ; preds =
%"164.i",
%"164.preheader.i"
%221 = phi i32 [ %236, %"164.i" ], [ 1, %"164.preheader.i"
]
%222 = sext i32 %221 to i64
%223 = add i64 %220, %222
%224 = getelementptr float* %206, i64 %223
%225 = load float* %224, align 4
%226 = call float @sinf(float %225) nounwind readnone
%227 = call float @asinf(float %226) nounwind readnone
%228 = add i64 %216, %222
%229 = getelementptr [0 x float]* %173, i64 0, i64 %228
%230 = load float* %229, align 4
%231 = call float @cosf(float %230) nounwind readnone
%232 = call float @acosf(float %231) nounwind readnone
%233 = fadd float %227, %232
%234 = getelementptr float* %208, i64 %223
store float %233, float* %234, align 4
%235 = icmp eq i32 %221, %.pre118.i
%236 = add i32 %221, 1
br i1 %235, label %"166.i", label %"164.i"
"166.i": ; preds =
%"164.i",
%"163.i"
%237 = icmp eq i32 %217, %.pre.i
%238 = add i32 %217, 1
br i1 %237, label %"168.i", label %"163.i"
"168.i": ; preds =
%"166.i",
%"162.i"
%239 = icmp eq i32 %210, %204
%240 = add i32 %210, 1
br i1 %239, label %"170.i", label %"162.i"
"170.i": ; preds =
%"168.i",
%"161.i"
call void bitcast (void (...)* @_gfortran_cpu_time_4 to void
(float*)*)(float* %finish.i) nounwind
br i1 %37, label %"172.i", label %"171.i"
GIMPLE:
======
_gfortran_cpu_time_4 (&start);
{
integer(kind=4) D.1699;
D.1699 = ns;
k = 1;
if (k <= D.1699) goto <D.2262>; else goto <D.2263>;
<D.2262>:
<D.2264>:
{
logical(kind=4) D.1710;
{
integer(kind=4) D.1702;
D.1702 = ny;
j = 1;
if (j <= D.1702) goto <D.2265>; else goto <D.2266>;
<D.2265>:
<D.2267>:
{
logical(kind=4) D.1709;
{
integer(kind=4) D.1705;
D.1705 = nx;
i = 1;
if (i <= D.1705) goto <D.2268>; else goto
<D.2269>;
<D.2268>:
<D.2270>:
{
logical(kind=4) D.1708;
D.2271 = xy.data;
D.2272 = (integer(kind=8)) i;
D.2273 = (integer(kind=8)) k;
D.2274 = xy.dim[2].stride;
D.2275 = D.2273 * D.2274;
D.2276 = (integer(kind=8)) j;
D.2277 = xy.dim[1].stride;
D.2278 = D.2276 * D.2277;
D.2279 = D.2275 + D.2278;
D.2280 = D.2272 + D.2279;
D.2281 = xy.offset;
D.2282 = D.2280 + D.2281;
D.2283 = x.data;
D.2284 = (integer(kind=8)) i;
D.2285 = (integer(kind=8)) k;
D.2286 = x.dim[2].stride;
D.2287 = D.2285 * D.2286;
D.2288 = (integer(kind=8)) j;
D.2289 = x.dim[1].stride;
D.2290 = D.2288 * D.2289;
D.2291 = D.2287 + D.2290;
D.2292 = D.2284 + D.2291;
D.2293 = x.offset;
D.2294 = D.2292 + D.2293;
D.2295 = MEM[(real(kind=4)[0:] *)D.2283][D.2294];
D.2296 = __builtin_sinf (D.2295);
D.2297 = __builtin_asinf (D.2296);
D.2298 = y.data;
D.2299 = (integer(kind=8)) i;
D.2300 = y.dim[2].stride;
D.2301 = y.dim[1].stride;
D.2302 = D.2300 + D.2301;
D.2303 = (integer(kind=8)) k;
D.2304 = D.2302 * D.2303;
D.2305 = D.2299 + D.2304;
D.2306 = y.offset;
D.2307 = D.2305 + D.2306;
D.2308 = MEM[(real(kind=4)[0:] *)D.2298][D.2307];
D.2309 = __builtin_cosf (D.2308);
D.2310 = __builtin_acosf (D.2309);
D.2311 = D.2297 + D.2310;
MEM[(real(kind=4)[0:] *)D.2271][D.2282] = D.2311;
L.18:
D.1708 = i == D.1705;
i = i + 1;
if (D.1708 != 0) goto L.19; else goto <D.2312>;
<D.2312>:
}
goto <D.2270>;
<D.2269>:
L.19:
}
L.16:
D.1709 = j == D.1702;
j = j + 1;
if (D.1709 != 0) goto L.17; else goto <D.2313>;
<D.2313>:
}
goto <D.2267>;
<D.2266>:
L.17:
}
L.14:
D.1710 = k == D.1699;
k = k + 1;
if (D.1710 != 0) goto L.15; else goto <D.2314>;
<D.2314>:
}
goto <D.2264>;
<D.2263>:
L.15:
}
_gfortran_cpu_time_4 (&finish);
2) Good LLVM IR:
===============
marcusmae at M17xR4:~/forge/kernelgen/tests/behavior/demo_f$
KERNELGEN_FALLBACK=1 kernelgen-gfortran -fplugin=dragonegg.so
-fplugin-arg-dragonegg-emit-ir -fplugin-arg-dragonegg-llvm-ir-optimize=0 -S
-c demo_f_m.f90 -o - | opt -O3 -S -o -
entry:
%0 = load i32* %nx, align 4
%1 = sext i32 %0 to i64
%2 = icmp slt i64 %1, 0
%3 = select i1 %2, i64 0, i64 %1
%4 = load i32* %ny, align 4
%5 = sext i32 %4 to i64
%6 = mul i64 %3, %5
%7 = icmp slt i64 %6, 0
%8 = select i1 %7, i64 0, i64 %6
%9 = load i32* %ns, align 4
%not = xor i64 %3, -1
%10 = sub i64 %not, %8
%11 = icmp sgt i32 %9, 0
br i1 %11, label %"3.preheader", label %return
"3.preheader": ; preds = %entry
%12 = icmp sgt i32 %4, 0
%13 = icmp sgt i32 %0, 0
%14 = add i64 %8, %3
br i1 %12, label %"4.preheader.us", label %return
"9.us": ; preds =
%"4.preheader.us",
%"7.us.us"
%15 = icmp eq i32 %17, %9
%16 = add i32 %17, 1
br i1 %15, label %return, label %"4.preheader.us"
"4.preheader.us": ; preds =
%"9.us",
%"3.preheader"
%17 = phi i32 [ %16, %"9.us" ], [ 1, %"3.preheader" ]
%18 = sext i32 %17 to i64
%19 = mul i64 %18, %8
%20 = add i64 %19, %10
%21 = mul i64 %18, %14
%22 = add i64 %21, %10
br i1 %13, label %"5.preheader.us.us", label %"9.us"
"7.us.us": ; preds =
%"5.us.us"
%23 = icmp eq i32 %25, %4
%24 = add i32 %25, 1
br i1 %23, label %"9.us", label %"5.preheader.us.us"
"5.preheader.us.us": ; preds =
%"4.preheader.us",
%"7.us.us"
%25 = phi i32 [ %24, %"7.us.us" ], [ 1, %"4.preheader.us"
]
%26 = sext i32 %25 to i64
%27 = mul i64 %26, %3
%28 = add i64 %20, %27
br label %"5.us.us"
"5.us.us": ; preds =
%"5.us.us", %"
5.preheader.us.us"
%29 = phi i32 [ %44, %"5.us.us" ], [ 1,
%"5.preheader.us.us" ]
%30 = sext i32 %29 to i64
%31 = add i64 %28, %30
%32 = getelementptr [0 x float]* %x, i64 0, i64 %31
%33 = load float* %32, align 4
%34 = tail call float @sinf(float %33) nounwind readnone
%35 = tail call float @asinf(float %34) nounwind readnone
%36 = add i64 %22, %30
%37 = getelementptr [0 x float]* %y, i64 0, i64 %36
%38 = load float* %37, align 4
%39 = tail call float @cosf(float %38) nounwind readnone
%40 = tail call float @acosf(float %39) nounwind readnone
%41 = fadd float %35, %40
%42 = getelementptr [0 x float]* %xy, i64 0, i64 %31
store float %41, float* %42, align 4
%43 = icmp eq i32 %29, %0
%44 = add i32 %29, 1
br i1 %43, label %"7.us.us", label %"5.us.us"
return: ; preds =
%"3.preheader",
%"9.us", %entry
ret void
GIMPLE:
======
D.2413 = *nx;
ubound.7 = (integer(kind=8)) D.2413;
stride.9 = ubound.7;
stride.9 = MAX_EXPR <stride.9, 0>;
D.2414 = *ny;
ubound.8 = (integer(kind=8)) D.2414;
stride.11 = stride.9 * ubound.8;
stride.11 = MAX_EXPR <stride.11, 0>;
D.2415 = *ns;
ubound.10 = (integer(kind=8)) D.2415;
size.13 = stride.11 * ubound.10;
size.13 = MAX_EXPR <size.13, 0>;
D.1583 = size.13 + -1;
size.127 = (bit_size_type) size.13;
D.1584 = size.127 * 32;
size.128 = (<unnamed-unsigned:64>) size.13;
D.1585 = size.128 * 4;
D.2418 = ~stride.9;
offset.12 = D.2418 - stride.11;
D.2419 = *nx;
ubound.0 = (integer(kind=8)) D.2419;
stride.2 = ubound.0;
stride.2 = MAX_EXPR <stride.2, 0>;
D.2420 = *ny;
ubound.1 = (integer(kind=8)) D.2420;
stride.4 = stride.2 * ubound.1;
stride.4 = MAX_EXPR <stride.4, 0>;
D.2421 = *ns;
ubound.3 = (integer(kind=8)) D.2421;
size.6 = stride.4 * ubound.3;
size.6 = MAX_EXPR <size.6, 0>;
D.1580 = size.6 + -1;
size.129 = (bit_size_type) size.6;
D.1581 = size.129 * 32;
size.130 = (<unnamed-unsigned:64>) size.6;
D.1582 = size.130 * 4;
D.2424 = ~stride.2;
offset.5 = D.2424 - stride.4;
D.2425 = *nx;
ubound.14 = (integer(kind=8)) D.2425;
stride.16 = ubound.14;
stride.16 = MAX_EXPR <stride.16, 0>;
D.2426 = *ny;
ubound.15 = (integer(kind=8)) D.2426;
stride.18 = stride.16 * ubound.15;
stride.18 = MAX_EXPR <stride.18, 0>;
D.2427 = *ns;
ubound.17 = (integer(kind=8)) D.2427;
size.20 = stride.18 * ubound.17;
size.20 = MAX_EXPR <size.20, 0>;
D.1577 = size.20 + -1;
size.131 = (bit_size_type) size.20;
D.1578 = size.131 * 32;
size.132 = (<unnamed-unsigned:64>) size.20;
D.1579 = size.132 * 4;
D.2430 = ~stride.16;
offset.19 = D.2430 - stride.18;
{
integer(kind=4) D.1565;
D.1565 = *ns;
k = 1;
if (k <= D.1565) goto <D.2431>; else goto <D.2432>;
<D.2431>:
<D.2433>:
{
logical(kind=4) D.1576;
{
integer(kind=4) D.1568;
D.1568 = *ny;
j = 1;
if (j <= D.1568) goto <D.2434>; else goto <D.2435>;
<D.2434>:
<D.2436>:
{
logical(kind=4) D.1575;
{
integer(kind=4) D.1571;
D.1571 = *nx;
i = 1;
if (i <= D.1571) goto <D.2437>; else goto <D.2438>;
<D.2437>:
<D.2439>:
{
logical(kind=4) D.1574;
D.2440 = (integer(kind=8)) i;
D.2441 = (integer(kind=8)) k;
D.2442 = D.2441 * stride.11;
D.2443 = (integer(kind=8)) j;
D.2444 = D.2443 * stride.9;
D.2445 = D.2442 + D.2444;
D.2446 = D.2440 + D.2445;
D.2447 = D.2446 + offset.12;
D.2448 = (integer(kind=8)) i;
D.2449 = (integer(kind=8)) k;
D.2450 = D.2449 * stride.4;
D.2451 = (integer(kind=8)) j;
D.2452 = D.2451 * stride.2;
D.2453 = D.2450 + D.2452;
D.2454 = D.2448 + D.2453;
D.2455 = D.2454 + offset.5;
D.2456 = *x[D.2455];
D.2457 = __builtin_sinf (D.2456);
D.2458 = __builtin_asinf (D.2457);
D.2459 = (integer(kind=8)) i;
D.2460 = stride.18 + stride.16;
D.2461 = (integer(kind=8)) k;
D.2462 = D.2460 * D.2461;
D.2463 = D.2459 + D.2462;
D.2464 = D.2463 + offset.19;
D.2465 = *y[D.2464];
D.2466 = __builtin_cosf (D.2465);
D.2467 = __builtin_acosf (D.2466);
D.2468 = D.2458 + D.2467;
*xy[D.2447] = D.2468;
L.5:
D.1574 = i == D.1571;
i = i + 1;
if (D.1574 != 0) goto L.6; else goto <D.2469>;
<D.2469>:
}
goto <D.2439>;
<D.2438>:
L.6:
}
L.3:
D.1575 = j == D.1568;
j = j + 1;
if (D.1575 != 0) goto L.4; else goto <D.2470>;
<D.2470>:
}
goto <D.2436>;
<D.2435>:
L.4:
}
L.1:
D.1576 = k == D.1565;
k = k + 1;
if (D.1576 != 0) goto L.2; else goto <D.2471>;
<D.2471>:
}
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20121231/92b218f2/attachment.html>
Dmitry Mikushin
2012-Dec-31 18:06 UTC
[LLVMdev] [DragonEgg] [Polly] Should we expect DragonEgg to produce identical LLVM IR for identical GIMPLE?
Forgot to attach original sources (Fortran). 2012/12/31 Dmitry Mikushin <dmitry at kernelgen.org>> Dear all, > > In our compiler we use a modified version LLVM Polly, which is very > sensitive to proper code generation. Among the number of limitations, the > loop region (enclosed by phi node on induction variable and branch) is > required to be free of additional memory-dependent branches. In other > words, there must be no conditional "br" instructions below phi nodes. The > problem we are facing is that from *identical* GIMPLE for 3d loop used in > different contexts DragonEgg may generate LLVM IR either conforming the > described limitation, or violating it. > > Let's consider two examples. In first one some simple 3D loop is used > directly in main program. In second - the same 3D loop is extracted into > separate subroutine called from main. Attached source code and listings for > GIMPLE and LLVM IR show that although GIMPLE codes are similar, by some > reason in first case branching goes under phi nodes, making Polly to fail > with parallelizing the resion. In second case everything is fine. > > I have not looked into DragonEgg internals enough yet, but before I'll do, > the question is: should we expect DragonEgg to produce identical LLVM IRs > for identical GIMPLEs? The case shown here is really really important. The > success of parallelization utilities that are currently in a quite good > shape (thanks to their developers!) nowadays heavily relies on the code > quality: if IR is too rough, we can't do much about it :( > > Many thanks and Happy New Year! > - Dima. > > > 1) Bad LLVM IR: > ==============> > marcusmae at M17xR4:~/forge/kernelgen/tests/behavior/demo_f$ > KERNELGEN_FALLBACK=1 kernelgen-gfortran -fplugin=dragonegg.so > -fplugin-arg-dragonegg-emit-ir -fplugin-arg-dragonegg-llvm-ir-optimize=0 -S > -c demo_f.f90 -o - | opt -O3 -S -o - > > "161.i": ; preds = %"160.i", > %"159.i" > call void bitcast (void (...)* @_gfortran_cpu_time_4 to void > (float*)*)(float* %start.i) nounwind > %204 = load i32* %ns.i, align 4 > %205 = icmp sgt i32 %204, 0 > br i1 %205, label %"162.preheader.i", label %"170.i" > > "162.preheader.i": ; preds = %"161.i" > %206 = bitcast i8* %x.0.0.i to float* > %207 = add i64 %y.3.2.0.0.i, %y.3.1.0.0.i > %208 = bitcast i8* %142 to float* > %.pre.i = load i32* %ny.i, align 4 > %209 = icmp sgt i32 %.pre.i, 0 > br label %"162.i" > > "162.i": ; preds = %"168.i", > %"162.preheader.i" > %210 = phi i32 [ %240, %"168.i" ], [ 1, %"162.preheader.i" ] > ; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ THIS BRANCH CREATES > A PROBLEM > br i1 %209, label %"163.preheader.i", label %"168.i" > ; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ THIS BRANCH CREATES > A PROBLEM > > "163.preheader.i": ; preds = %"162.i" > %211 = sext i32 %210 to i64 > %212 = mul i64 %211, %86 > %213 = mul i64 %211, %207 > %.pre118.i = load i32* %nx.i, align 4 > %214 = icmp sgt i32 %.pre118.i, 0 > %215 = add i64 %212, %98 > %216 = add i64 %213, %y.1.0.i > br label %"163.i" > > "163.i": ; preds = %"166.i", > %"163.preheader.i" > %217 = phi i32 [ %238, %"166.i" ], [ 1, %"163.preheader.i" ] > ; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ THIS BRANCH CREATES > A PROBLEM > br i1 %214, label %"164.preheader.i", label %"166.i" > ; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ THIS BRANCH CREATES > A PROBLEM > > "164.preheader.i": ; preds = %"163.i" > %218 = sext i32 %217 to i64 > %219 = mul i64 %218, %72 > %220 = add i64 %215, %219 > br label %"164.i" > > "164.i": ; preds = %"164.i", > %"164.preheader.i" > %221 = phi i32 [ %236, %"164.i" ], [ 1, %"164.preheader.i" ] > %222 = sext i32 %221 to i64 > %223 = add i64 %220, %222 > %224 = getelementptr float* %206, i64 %223 > %225 = load float* %224, align 4 > %226 = call float @sinf(float %225) nounwind readnone > %227 = call float @asinf(float %226) nounwind readnone > %228 = add i64 %216, %222 > %229 = getelementptr [0 x float]* %173, i64 0, i64 %228 > %230 = load float* %229, align 4 > %231 = call float @cosf(float %230) nounwind readnone > %232 = call float @acosf(float %231) nounwind readnone > %233 = fadd float %227, %232 > %234 = getelementptr float* %208, i64 %223 > store float %233, float* %234, align 4 > %235 = icmp eq i32 %221, %.pre118.i > %236 = add i32 %221, 1 > br i1 %235, label %"166.i", label %"164.i" > > "166.i": ; preds = %"164.i", > %"163.i" > %237 = icmp eq i32 %217, %.pre.i > %238 = add i32 %217, 1 > br i1 %237, label %"168.i", label %"163.i" > > "168.i": ; preds = %"166.i", > %"162.i" > %239 = icmp eq i32 %210, %204 > %240 = add i32 %210, 1 > br i1 %239, label %"170.i", label %"162.i" > > "170.i": ; preds = %"168.i", > %"161.i" > call void bitcast (void (...)* @_gfortran_cpu_time_4 to void > (float*)*)(float* %finish.i) nounwind > br i1 %37, label %"172.i", label %"171.i" > > GIMPLE: > ======> > _gfortran_cpu_time_4 (&start); > { > integer(kind=4) D.1699; > > D.1699 = ns; > k = 1; > if (k <= D.1699) goto <D.2262>; else goto <D.2263>; > <D.2262>: > <D.2264>: > { > logical(kind=4) D.1710; > > { > integer(kind=4) D.1702; > > D.1702 = ny; > j = 1; > if (j <= D.1702) goto <D.2265>; else goto <D.2266>; > <D.2265>: > <D.2267>: > { > logical(kind=4) D.1709; > > { > integer(kind=4) D.1705; > > D.1705 = nx; > i = 1; > if (i <= D.1705) goto <D.2268>; else goto <D.2269>; > <D.2268>: > <D.2270>: > { > logical(kind=4) D.1708; > > D.2271 = xy.data; > D.2272 = (integer(kind=8)) i; > D.2273 = (integer(kind=8)) k; > D.2274 = xy.dim[2].stride; > D.2275 = D.2273 * D.2274; > D.2276 = (integer(kind=8)) j; > D.2277 = xy.dim[1].stride; > D.2278 = D.2276 * D.2277; > D.2279 = D.2275 + D.2278; > D.2280 = D.2272 + D.2279; > D.2281 = xy.offset; > D.2282 = D.2280 + D.2281; > D.2283 = x.data; > D.2284 = (integer(kind=8)) i; > D.2285 = (integer(kind=8)) k; > D.2286 = x.dim[2].stride; > D.2287 = D.2285 * D.2286; > D.2288 = (integer(kind=8)) j; > D.2289 = x.dim[1].stride; > D.2290 = D.2288 * D.2289; > D.2291 = D.2287 + D.2290; > D.2292 = D.2284 + D.2291; > D.2293 = x.offset; > D.2294 = D.2292 + D.2293; > D.2295 = MEM[(real(kind=4)[0:] *)D.2283][D.2294]; > D.2296 = __builtin_sinf (D.2295); > D.2297 = __builtin_asinf (D.2296); > D.2298 = y.data; > D.2299 = (integer(kind=8)) i; > D.2300 = y.dim[2].stride; > D.2301 = y.dim[1].stride; > D.2302 = D.2300 + D.2301; > D.2303 = (integer(kind=8)) k; > D.2304 = D.2302 * D.2303; > D.2305 = D.2299 + D.2304; > D.2306 = y.offset; > D.2307 = D.2305 + D.2306; > D.2308 = MEM[(real(kind=4)[0:] *)D.2298][D.2307]; > D.2309 = __builtin_cosf (D.2308); > D.2310 = __builtin_acosf (D.2309); > D.2311 = D.2297 + D.2310; > MEM[(real(kind=4)[0:] *)D.2271][D.2282] = D.2311; > L.18: > D.1708 = i == D.1705; > i = i + 1; > if (D.1708 != 0) goto L.19; else goto <D.2312>; > <D.2312>: > } > goto <D.2270>; > <D.2269>: > L.19: > } > L.16: > D.1709 = j == D.1702; > j = j + 1; > if (D.1709 != 0) goto L.17; else goto <D.2313>; > <D.2313>: > } > goto <D.2267>; > <D.2266>: > L.17: > } > L.14: > D.1710 = k == D.1699; > k = k + 1; > if (D.1710 != 0) goto L.15; else goto <D.2314>; > <D.2314>: > } > goto <D.2264>; > <D.2263>: > L.15: > } > _gfortran_cpu_time_4 (&finish); > > 2) Good LLVM IR: > ===============> > > marcusmae at M17xR4:~/forge/kernelgen/tests/behavior/demo_f$ > KERNELGEN_FALLBACK=1 kernelgen-gfortran -fplugin=dragonegg.so > -fplugin-arg-dragonegg-emit-ir -fplugin-arg-dragonegg-llvm-ir-optimize=0 -S > -c demo_f_m.f90 -o - | opt -O3 -S -o - > > entry: > %0 = load i32* %nx, align 4 > %1 = sext i32 %0 to i64 > %2 = icmp slt i64 %1, 0 > %3 = select i1 %2, i64 0, i64 %1 > %4 = load i32* %ny, align 4 > %5 = sext i32 %4 to i64 > %6 = mul i64 %3, %5 > %7 = icmp slt i64 %6, 0 > %8 = select i1 %7, i64 0, i64 %6 > %9 = load i32* %ns, align 4 > %not = xor i64 %3, -1 > %10 = sub i64 %not, %8 > %11 = icmp sgt i32 %9, 0 > br i1 %11, label %"3.preheader", label %return > > "3.preheader": ; preds = %entry > %12 = icmp sgt i32 %4, 0 > %13 = icmp sgt i32 %0, 0 > %14 = add i64 %8, %3 > br i1 %12, label %"4.preheader.us", label %return > > "9.us": ; preds = %" > 4.preheader.us", %"7.us.us" > %15 = icmp eq i32 %17, %9 > %16 = add i32 %17, 1 > br i1 %15, label %return, label %"4.preheader.us" > > "4.preheader.us": ; preds = %"9.us", > %"3.preheader" > %17 = phi i32 [ %16, %"9.us" ], [ 1, %"3.preheader" ] > %18 = sext i32 %17 to i64 > %19 = mul i64 %18, %8 > %20 = add i64 %19, %10 > %21 = mul i64 %18, %14 > %22 = add i64 %21, %10 > br i1 %13, label %"5.preheader.us.us", label %"9.us" > > "7.us.us": ; preds = %"5.us.us" > %23 = icmp eq i32 %25, %4 > %24 = add i32 %25, 1 > br i1 %23, label %"9.us", label %"5.preheader.us.us" > > "5.preheader.us.us": ; preds = %" > 4.preheader.us", %"7.us.us" > %25 = phi i32 [ %24, %"7.us.us" ], [ 1, %"4.preheader.us" ] > %26 = sext i32 %25 to i64 > %27 = mul i64 %26, %3 > %28 = add i64 %20, %27 > br label %"5.us.us" > > "5.us.us": ; preds = %"5.us.us", %" > 5.preheader.us.us" > %29 = phi i32 [ %44, %"5.us.us" ], [ 1, %"5.preheader.us.us" ] > %30 = sext i32 %29 to i64 > %31 = add i64 %28, %30 > %32 = getelementptr [0 x float]* %x, i64 0, i64 %31 > %33 = load float* %32, align 4 > %34 = tail call float @sinf(float %33) nounwind readnone > %35 = tail call float @asinf(float %34) nounwind readnone > %36 = add i64 %22, %30 > %37 = getelementptr [0 x float]* %y, i64 0, i64 %36 > %38 = load float* %37, align 4 > %39 = tail call float @cosf(float %38) nounwind readnone > %40 = tail call float @acosf(float %39) nounwind readnone > %41 = fadd float %35, %40 > %42 = getelementptr [0 x float]* %xy, i64 0, i64 %31 > store float %41, float* %42, align 4 > %43 = icmp eq i32 %29, %0 > %44 = add i32 %29, 1 > br i1 %43, label %"7.us.us", label %"5.us.us" > > return: ; preds > %"3.preheader", %"9.us", %entry > ret void > > GIMPLE: > ======> > D.2413 = *nx; > ubound.7 = (integer(kind=8)) D.2413; > stride.9 = ubound.7; > stride.9 = MAX_EXPR <stride.9, 0>; > D.2414 = *ny; > ubound.8 = (integer(kind=8)) D.2414; > stride.11 = stride.9 * ubound.8; > stride.11 = MAX_EXPR <stride.11, 0>; > D.2415 = *ns; > ubound.10 = (integer(kind=8)) D.2415; > size.13 = stride.11 * ubound.10; > size.13 = MAX_EXPR <size.13, 0>; > D.1583 = size.13 + -1; > size.127 = (bit_size_type) size.13; > D.1584 = size.127 * 32; > size.128 = (<unnamed-unsigned:64>) size.13; > D.1585 = size.128 * 4; > D.2418 = ~stride.9; > offset.12 = D.2418 - stride.11; > D.2419 = *nx; > ubound.0 = (integer(kind=8)) D.2419; > stride.2 = ubound.0; > stride.2 = MAX_EXPR <stride.2, 0>; > D.2420 = *ny; > ubound.1 = (integer(kind=8)) D.2420; > stride.4 = stride.2 * ubound.1; > stride.4 = MAX_EXPR <stride.4, 0>; > D.2421 = *ns; > ubound.3 = (integer(kind=8)) D.2421; > size.6 = stride.4 * ubound.3; > size.6 = MAX_EXPR <size.6, 0>; > D.1580 = size.6 + -1; > size.129 = (bit_size_type) size.6; > D.1581 = size.129 * 32; > size.130 = (<unnamed-unsigned:64>) size.6; > D.1582 = size.130 * 4; > D.2424 = ~stride.2; > offset.5 = D.2424 - stride.4; > D.2425 = *nx; > ubound.14 = (integer(kind=8)) D.2425; > stride.16 = ubound.14; > stride.16 = MAX_EXPR <stride.16, 0>; > D.2426 = *ny; > ubound.15 = (integer(kind=8)) D.2426; > stride.18 = stride.16 * ubound.15; > stride.18 = MAX_EXPR <stride.18, 0>; > D.2427 = *ns; > ubound.17 = (integer(kind=8)) D.2427; > size.20 = stride.18 * ubound.17; > size.20 = MAX_EXPR <size.20, 0>; > D.1577 = size.20 + -1; > size.131 = (bit_size_type) size.20; > D.1578 = size.131 * 32; > size.132 = (<unnamed-unsigned:64>) size.20; > D.1579 = size.132 * 4; > D.2430 = ~stride.16; > offset.19 = D.2430 - stride.18; > { > integer(kind=4) D.1565; > > D.1565 = *ns; > k = 1; > if (k <= D.1565) goto <D.2431>; else goto <D.2432>; > <D.2431>: > <D.2433>: > { > logical(kind=4) D.1576; > > { > integer(kind=4) D.1568; > > D.1568 = *ny; > j = 1; > if (j <= D.1568) goto <D.2434>; else goto <D.2435>; > <D.2434>: > <D.2436>: > { > logical(kind=4) D.1575; > > { > integer(kind=4) D.1571; > > D.1571 = *nx; > i = 1; > if (i <= D.1571) goto <D.2437>; else goto <D.2438>; > <D.2437>: > <D.2439>: > { > logical(kind=4) D.1574; > > D.2440 = (integer(kind=8)) i; > D.2441 = (integer(kind=8)) k; > D.2442 = D.2441 * stride.11; > D.2443 = (integer(kind=8)) j; > D.2444 = D.2443 * stride.9; > D.2445 = D.2442 + D.2444; > D.2446 = D.2440 + D.2445; > D.2447 = D.2446 + offset.12; > D.2448 = (integer(kind=8)) i; > D.2449 = (integer(kind=8)) k; > D.2450 = D.2449 * stride.4; > D.2451 = (integer(kind=8)) j; > D.2452 = D.2451 * stride.2; > D.2453 = D.2450 + D.2452; > D.2454 = D.2448 + D.2453; > D.2455 = D.2454 + offset.5; > D.2456 = *x[D.2455]; > D.2457 = __builtin_sinf (D.2456); > D.2458 = __builtin_asinf (D.2457); > D.2459 = (integer(kind=8)) i; > D.2460 = stride.18 + stride.16; > D.2461 = (integer(kind=8)) k; > D.2462 = D.2460 * D.2461; > D.2463 = D.2459 + D.2462; > D.2464 = D.2463 + offset.19; > D.2465 = *y[D.2464]; > D.2466 = __builtin_cosf (D.2465); > D.2467 = __builtin_acosf (D.2466); > D.2468 = D.2458 + D.2467; > *xy[D.2447] = D.2468; > L.5: > D.1574 = i == D.1571; > i = i + 1; > if (D.1574 != 0) goto L.6; else goto <D.2469>; > <D.2469>: > } > goto <D.2439>; > <D.2438>: > L.6: > } > L.3: > D.1575 = j == D.1568; > j = j + 1; > if (D.1575 != 0) goto L.4; else goto <D.2470>; > <D.2470>: > } > goto <D.2436>; > <D.2435>: > L.4: > } > L.1: > D.1576 = k == D.1565; > k = k + 1; > if (D.1576 != 0) goto L.2; else goto <D.2471>; > <D.2471>: > } > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20121231/1b365ef3/attachment.html> -------------- next part -------------- A non-text attachment was scrubbed... Name: demo_f.f90 Type: application/octet-stream Size: 1524 bytes Desc: not available URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20121231/1b365ef3/attachment.obj> -------------- next part -------------- A non-text attachment was scrubbed... Name: demo_f_m.f90 Type: application/octet-stream Size: 1710 bytes Desc: not available URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20121231/1b365ef3/attachment-0001.obj>
Duncan Sands
2013-Jan-01 13:45 UTC
[LLVMdev] [DragonEgg] [Polly] Should we expect DragonEgg to produce identical LLVM IR for identical GIMPLE?
Hi Dmitry,> > In our compiler we use a modified version LLVM Polly, which is very sensitive to > proper code generation. Among the number of limitations, the loop region > (enclosed by phi node on induction variable and branch) is required to be free > of additional memory-dependent branches. In other words, there must be no > conditional "br" instructions below phi nodes. The problem we are facing is that > from *identical* GIMPLE for 3d loop used in different contexts DragonEgg may > generate LLVM IR either conforming the described limitation, or violating it.the gimple isn't the same at all (see below). The differences are directly reflected in the unoptimized LLVM IR, turning up as additional memory loads in the "bad" version. In addition, the Fortran isn't really the same either: Fortran semantics allows the compiler to assume that the parameters of your new function "compute" (which are all passed in by reference, i.e. as pointers) do not alias each other or anything else in sight (i.e. they get the "restrict" qualifier in the gimple, noalias in the LLVM IR). Thus by factorizing the loop into "compute" you are actually giving the compiler more information. Summary: (1) as far as I can see the unoptimized LLVM IR is a direct reflection of the gimple: the differences for the loop part come directly from differences in the gimple; (2) the optimizers do a better good when you have "compute" partly because you provided them with additional aliasing information; this better optimized version then gets inlined into MAIN__. (3) this leaves the question of whether in the bad version it is logically possible for the optimizers to deduce the same aliasing information as is handed to them for free in the good version. To work this out it would be nice to have a smaller testcase. Ciao, Duncan. PS: Gimple differences for the loop part. I replaced all variable indices and such with X otherwise they get in the way of the diffing. <bb X>: # k_X = PHI <k_X(X), k_X(X)> - D.X_X = ny; + D.X_X = *ny_X(D); j_X = X; if (j_X <= D.X_X) goto <bb X>; @@ -16,7 +74,7 @@ <bb X>: # j_X = PHI <j_X(X), j_X(X)> - D.X_X = nx; + D.X_X = *nx_X(D); i_X = X; if (i_X <= D.X_X) goto <bb X>; @@ -25,48 +83,36 @@ <bb X>: # i_X = PHI <i_X(X), i_X(X)> - D.X_X = xy.data; D.X_X = (integer(kind=X)) i_X; D.X_X = (integer(kind=X)) k_X; - D.X_X = xy.dim[X].stride; - D.X_X = D.X_X * D.X_X; + D.X_X = D.X_X * stride.X_X; D.X_X = (integer(kind=X)) j_X; - D.X_X = xy.dim[X].stride; - D.X_X = D.X_X * D.X_X; + D.X_X = D.X_X * stride.X_X; D.X_X = D.X_X + D.X_X; D.X_X = D.X_X + D.X_X; - D.X_X = xy.offset; - D.X_X = D.X_X + D.X_X; - D.X_X = x.data; + D.X_X = D.X_X + offset.X_X; D.X_X = (integer(kind=X)) i_X; D.X_X = (integer(kind=X)) k_X; - D.X_X = x.dim[X].stride; - D.X_X = D.X_X * D.X_X; + D.X_X = D.X_X * stride.X_X; D.X_X = (integer(kind=X)) j_X; - D.X_X = x.dim[X].stride; - D.X_X = D.X_X * D.X_X; + D.X_X = D.X_X * stride.X_X; D.X_X = D.X_X + D.X_X; D.X_X = D.X_X + D.X_X; - D.X_X = x.offset; - D.X_X = D.X_X + D.X_X; - D.X_X = MEM[(real(kind=X)[X:] *)D.X_X][D.X_X]; + D.X_X = D.X_X + offset.X_X; + D.X_X = *x_X(D)[D.X_X]; D.X_X = sinf (D.X_X); D.X_X = asinf (D.X_X); - D.X_X = y.data; D.X_X = (integer(kind=X)) i_X; - D.X_X = y.dim[X].stride; - D.X_X = y.dim[X].stride; - D.X_X = D.X_X + D.X_X; + D.X_X = stride.X_X + stride.X_X; D.X_X = (integer(kind=X)) k_X; D.X_X = D.X_X * D.X_X; D.X_X = D.X_X + D.X_X; - D.X_X = y.offset; - D.X_X = D.X_X + D.X_X; - D.X_X = MEM[(real(kind=X)[X:] *)D.X_X][D.X_X]; + D.X_X = D.X_X + offset.X_X; + D.X_X = *y_X(D)[D.X_X]; D.X_X = cosf (D.X_X); D.X_X = acosf (D.X_X); D.X_X = D.X_X + D.X_X; - MEM[(real(kind=X)[X:] *)D.X_X][D.X_X] = D.X_X; + *xy_X(D)[D.X_X] = D.X_X; D.X_X = i_X == D.X_X; i_X = i_X + X; if (D.X_X != X)
Tobias Grosser
2013-Jan-02 11:02 UTC
[LLVMdev] [DragonEgg] [Polly] Should we expect DragonEgg to produce identical LLVM IR for identical GIMPLE?
On 01/01/2013 02:45 PM, Duncan Sands wrote:> Hi Dmitry, > >> >> In our compiler we use a modified version LLVM Polly, which is very >> sensitive to >> proper code generation. Among the number of limitations, the loop region >> (enclosed by phi node on induction variable and branch) is required to >> be free >> of additional memory-dependent branches. In other words, there must be no >> conditional "br" instructions below phi nodes. The problem we are >> facing is that >> from *identical* GIMPLE for 3d loop used in different contexts >> DragonEgg may >> generate LLVM IR either conforming the described limitation, or >> violating it. > > the gimple isn't the same at all (see below). The differences are directly > reflected in the unoptimized LLVM IR, turning up as additional memory loads > in the "bad" version. In addition, the Fortran isn't really the same > either: > Fortran semantics allows the compiler to assume that the parameters of your > new function "compute" (which are all passed in by reference, i.e. as > pointers) > do not alias each other or anything else in sight (i.e. they get the > "restrict" > qualifier in the gimple, noalias in the LLVM IR). Thus by factorizing > the loop > into "compute" you are actually giving the compiler more information. > > Summary: > (1) as far as I can see the unoptimized LLVM IR is a direct > reflection of > the gimple: the differences for the loop part come directly from > differences > in the gimple; > (2) the optimizers do a better good when you have "compute" partly > because you > provided them with additional aliasing information; this better optimized > version then gets inlined into MAIN__. > (3) this leaves the question of whether in the bad version it is > logically > possible for the optimizers to deduce the same aliasing information as is > handed to them for free in the good version. To work this out it would be > nice to have a smaller testcase.I would also be interested in a minimal test case. If e.g. only the alias check is missing, we could introduce run-time alias checks such that Polly would be able to optimize both versions. It is probably not as simple, but a reduced test case would make it easier to figure out the exact problems. Thanks Tobi
Reasonably Related Threads
- [LLVMdev] [DragonEgg] [Polly] Should we expect DragonEgg to produce identical LLVM IR for identical GIMPLE?
- [LLVMdev] [Polly] Aliasing problems escalation (WAS: Re: [DragonEgg] [Polly] Should we expect DragonEgg to produce identical LLVM IR for identical GIMPLE?)
- How to optimize store of constant arrays
- [LLVMdev] reference to non-static member function must be called
- [LLVMdev] How to deal with wider Integer type?